Overclock.net › Forums › Industry News › Software News › [phoronix] Cycles patch enables 400% speed up on AMD cards!
New Posts  All Forums:Forum Nav:

[phoronix] Cycles patch enables 400% speed up on AMD cards!

post #1 of 17
Thread Starter 

Quote:
George Kyriazis of AMD has provided patches to the Blender project for vastly improving their OpenCL Cycles renderer support and allow for it to work with AMD GPUs.

Kyriazis wrote a Blender patch for splitting the Cycles' OpenCL mega-kernel into separate kernels. By breaking down the big compute kernel, better GPU utilization can be obtained and better performance. While George posted the patch, other AMD developers were also involved in redoing this OpenCL code for the Cycles renderer.

User benchmarks are showing huge improvements in the performance and also a slight reduction in CPU usage. While the work is still being reviewed, the patch can be viewed via developer.blender.org.
1. User run benchmarks
2. Source
3. AMD initiative last year to fix GPU issues for both Radeon and GeForce
post #2 of 17
Spoiler: GCN and newer.

Also, don't expect it to be in main until much later.

And on a side note with this patch the 290X is as fast as a GTX 580 and both are rolfstomped by a GTX 970/80.
post #3 of 17
Thread Starter 
Quote:
Spoiler: GCN and newer.

Also, don't expect it to be in main until much later.

And on a side note with this patch the 290X is as fast as a GTX 580 and both are rolfstomped by a GTX 970/80.
Yeah, it'll take a while to land in the release branch.
But actually, according to #2023, for new BMW benchmark the 290x using AMD's new build is nearly as fast as the 780 using CUDA, 88 seconds vs 82 seconds.
edit: grammar
edit 2: Thanzex indicates the 770 takes 110 seconds to render the BMW new bench, although it's not clear if he's using CUDA or opencl.
Edited by SandGlass - 3/28/15 at 4:29pm
post #4 of 17
Nope.

There are several versions of the BMW bench. In the V4 I score exactly the same as this user with my stock GTX 580 3GB.

I couldn't care less if new hardware is faster, just pointing out the only comparable data I found.
post #5 of 17
Thread Starter 
The most recent BMW version is the one updated 2 months ago, with 2 cars instead of one. Your results are pretty interesting, are you using 2.73a or 2.74 RC2? I'm curious because your results would indicate that it rivals the titan depending on the cycles version.

Edit: Never mind, found multiple sources indicating that the Nvidia's performance for the 770 is around that of the 580.
Edited by SandGlass - 3/28/15 at 4:49pm
post #6 of 17
These benches have the version in a little txt, you don't have to guess based on its release.

UuRjMcc.png

Blender 2.74RC2
BMW V4
Linux Mint 17

If you check the link in my previous you'll see that user's 290X doing it just 10 secs faster.

EDIT: And no, the GTX 770 is slower than the GTX 580.

EDIT 2: Some data gathered a while back: https://docs.google.com/spreadsheet/pub?key=0AuyaBndR19VBdEdLVzI0c0ZwV3NyRXU1amk3aHhJaXc&output=html

And some more: http://blenchmark.com/gpu-benchmarks
Edited by Imouto - 3/28/15 at 5:15pm
post #7 of 17
Thread Starter 
That's pretty cool! Looking at the spread sheet of the results(old) there seems to be very little correlation between speed and CUDA core count, instead VRAM bandwidth seems to be the most important factor. Whether or not this is true for AMD cards is unknown since I don't run one, and there aren't enough results online right now.

Edit: I was looking at the BMW benchmark sheet
Edit 2: The performance index on the right seems to say the 580 3GB blows the 770 out of the water.

Edited by SandGlass - 3/28/15 at 5:23pm
post #8 of 17
All of them have their pros and cons. The GTX 770 is faster in viewport perf, meaning that it is more pleasant to model and do stuff.

I Haven't checked in ages but AMD always been better in that regard because of LUTs, double faced texturing and being better at OpenGL overall.

So AMD doing better with Blender may make me look at one of their cards in years.
post #9 of 17
What's viewport/redraw accuracy like with current AMD GPU's/drivers?
post #10 of 17
Could this be done in all OpenCL applications? If so does that mean GPU mining could become profitable again? Or if not with R9 200 series possibly R9 300 series?
New Posts  All Forums:Forum Nav:
  Return Home
  Back to Forum: Software News
Overclock.net › Forums › Industry News › Software News › [phoronix] Cycles patch enables 400% speed up on AMD cards!